CN102313897A

CN102313897A - Radioactive spectrum identification method

Info

Publication number: CN102313897A
Application number: CN2010102123506A
Authority: CN
Inventors: 黄洪全; 方方; 阎萍; 王超; 王敏; 龚迪琛; 丁卫撑; 刘念聪; 周伟; 刘易
Original assignee: Chengdu Univeristy of Technology
Current assignee: Chengdu Univeristy of Technology
Priority date: 2010-06-29
Filing date: 2010-06-29
Publication date: 2012-01-11

Abstract

The invention discloses a radioactivity spectrum identification method. Perform filtering, dimension reduction and classification processing on the measured energy spectrum in radioactivity measurement, extract its feature vector as a sample, and obtain the GMM model of various feature vectors through training; find the feature vector of the energy spectrum to be identified, and press GMM The model obtains the class conditional probability density of the feature vector, and then decides to classify, that is, completes the energy spectrum recognition. The invention has high recognition accuracy for energy spectrum, and is an effective method for identifying radioactive energy spectrum.

Description

A radioactivity spectrum identification method

技术领域 technical field

本发明涉及一种放射性能谱识别方法。The invention relates to a radioactivity spectrum identification method.

背景技术 Background technique

在进行放射性的能谱测量中，常常要根据探测器采集到的大量随机数据，分析这些随机信号由什么放射源产生。一般的放射性能谱解析方法是对所测的能谱在进行了平滑滤波、扣除本底、寻峰、计算净峰面积等分析后，根据能量、峰形和效率刻度的结果，由峰位所对应的能量识别出放射源含有哪些放射性核素，进而由净峰面积计算出这些放射性核素的活度。在某些应用场合，由于核信息保护的原因，不需要得到核素活度的精确结果，只需识别所测对象为何种核材料。为此，曾有研究者采用能谱信息作为神经网络的训练样本，并用训练后的神经网络对这些核材料进行识别。这些方法存在难以训练、或者不易收敛、或者收敛到局部极小等缺点，导致能谱的识别不能进行或者识别准确率降低，故需要找出一种既能快速训练又能避免不能收敛或局部收敛的方法，使能谱的识别能顺利进行且能保证更高的识别准确率。In the energy spectrum measurement of radioactivity, it is often necessary to analyze what radioactive source these random signals are produced based on a large amount of random data collected by the detector. The general radioactivity spectrum analysis method is to analyze the measured energy spectrum by smoothing, filtering, background subtraction, peak finding, and calculating the net peak area, etc., according to the results of the energy, peak shape and efficiency scale, the peak position is determined. The corresponding energies identify which radionuclides the source contains, and the activities of these radionuclides are calculated from the net peak areas. In some applications, due to the protection of nuclear information, it is not necessary to obtain accurate results of nuclide activity, but only to identify what kind of nuclear material the measured object is. For this reason, some researchers used the energy spectrum information as the training samples of the neural network, and used the trained neural network to identify these nuclear materials. These methods have disadvantages such as difficult to train, or difficult to converge, or converge to a local minimum, which leads to the inability to recognize the energy spectrum or reduce the recognition accuracy. The method can make the recognition of energy spectrum go smoothly and can guarantee higher recognition accuracy.

发明内容 Contents of the invention

本发明的目的在于公开一种放射性能谱识别方法。该方法克服了现有能谱识别方法的不足，具有训练速度快、收敛准确、识别准确率高等特点，是一种进行放射性能谱识别的有效方法。The purpose of the present invention is to disclose a radioactivity spectrum identification method. This method overcomes the shortcomings of existing energy spectrum identification methods, and has the characteristics of fast training speed, accurate convergence, and high recognition accuracy, and is an effective method for radioactivity energy spectrum identification.

本发明是通过以下技术方案实现的，本发明的具体步骤如下：The present invention is achieved through the following technical solutions, and the concrete steps of the present invention are as follows:

①对放射性测量中已测得的能谱进行训练，按如下A～E步骤：① To train the energy spectrum measured in the radioactivity measurement, follow the steps A~E as follows:

A根据需要按放射源类别对已测得的多个能谱进行分类，A Classify the measured multiple energy spectra according to the category of radioactive sources as needed,

B采用小波方法或多项式方法对能谱进行平滑滤波，B Use wavelet method or polynomial method to smooth and filter the energy spectrum,

C求取能谱的GMM模型(Gaussian mixture model)，并将GMM模型中各高斯函数的权值作为降维后的数据，C obtains the GMM model (Gaussian mixture model) of the energy spectrum, and uses the weight of each Gaussian function in the GMM model as the data after dimension reduction,

D将降维后的能谱数据进行小波包分解，并将各子频带信号的能量进行归一化处理，提取归一化数据作为特征向量。D Decompose the dimensionally reduced energy spectrum data into wavelet packets, normalize the energy of each sub-band signal, and extract the normalized data as feature vectors.

E将特征向量作为样本，采用期望最大化法(Expectation Maximization，简写为EM)对这些样本进行迭代运算，得到各类特征向量的GMM模型；E takes the eigenvectors as samples, and uses the Expectation Maximization method (Expectation Maximization, abbreviated as EM) to perform iterative operations on these samples to obtain the GMM model of various eigenvectors;

②将待识别能谱进行决策分类，按如下A～E步骤：② Decision-making classification of the energy spectrum to be identified, according to the following steps A~E:

A采用小波方法或多项式方法对待识别能谱进行平滑滤波，A Use wavelet method or polynomial method to smooth filter the energy spectrum to be identified,

B求取待识别能谱的GMM模型(Gaussian mixture model)，并将GMM模型中各高斯函数的权值作为降维后的数据，B Find the GMM model (Gaussian mixture model) of the energy spectrum to be identified, and use the weight of each Gaussian function in the GMM model as the data after dimension reduction,

C将降维后的待识别能谱数据进行小波包分解，并将各子频带信号的能量进行归一化处理，提取归一化数据作为特征向量，C decomposes the energy spectrum data to be identified after dimension reduction by wavelet packet, and normalizes the energy of each sub-band signal, and extracts the normalized data as the feature vector,

D将待识别能谱的特征向量作为多维随机数，并计算其分属于各类GMM模型的类条件概率密度，D uses the eigenvector of the energy spectrum to be identified as a multi-dimensional random number, and calculates the class conditional probability density belonging to each type of GMM model,

E最后按贝叶斯决策分类。E is finally classified by Bayesian decision.

本发明的有益效果是：The beneficial effects of the present invention are:

本发明在训练阶段先采用小波方法或多项式方法对能谱进行平滑滤波，以消除干扰信号；接着用GMM模型对能谱作降维处理，以保证能谱的统计特性在低维空间仍具有原始能谱信号的绝大部分信息；然后，在多频带信号空间提取归一化特征向量并作为样本，保证了采用低维样本代表能谱特征的唯一性，同时也保证了样本与测量时间的无关性；最后，采用期望最大化法对这些样本进行迭代运算，得到各类特征向量的GMM模型，这样既保证了准确收敛又可通过合理调整GMM模型的高斯函数个数以满足各种不同场合下识别的需要。另外，本发明在识别时，采用贝叶斯决策分类，保证了识别准确率在统计意义下的最优。总之，本发明是一种既能快速训练又能避免不能收敛或局部收敛的方法，使能谱的识别能顺利进行且能保证更高的识别准确率。In the training stage, the present invention first adopts the wavelet method or polynomial method to smooth and filter the energy spectrum to eliminate interference signals; then uses the GMM model to perform dimensionality reduction processing on the energy spectrum to ensure that the statistical characteristics of the energy spectrum still have the original Most of the information of the energy spectrum signal; then, the normalized feature vector is extracted in the multi-band signal space and used as a sample, which ensures the uniqueness of using low-dimensional samples to represent the energy spectrum features, and also ensures that the sample has nothing to do with the measurement time Finally, the expectation maximization method is used to iteratively operate these samples to obtain the GMM model of various eigenvectors. identification needs. In addition, the present invention adopts Bayesian decision-making classification during recognition, which ensures the optimal recognition accuracy rate in a statistical sense. In a word, the present invention is a method capable of fast training and avoiding non-convergence or partial convergence, so that the identification of energy spectrum can be carried out smoothly and can ensure higher identification accuracy.

附图说明 Description of drawings

图1为本发明方法的流程图；Fig. 1 is the flowchart of the inventive method;

具体实施方式 Detailed ways

为使发明的目的、技术方案及优点更加清楚明白，以下参照附图并举实施例，对本发明作进一步详细说明。In order to make the object, technical solution and advantages of the invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples.

由于核信息保护的原因，不需要得到核素活度精确结果，只需识别所测对象为何种核材料，针对这一情况，本发明提供了一种放射性能谱识别方法。图1显示了本发明所述识别方法的流程。Due to the protection of nuclear information, it is not necessary to obtain accurate results of nuclide activity, but only to identify what kind of nuclear material the measured object is. Aiming at this situation, the present invention provides a radioactivity spectrum identification method. Fig. 1 shows the flow chart of the identification method of the present invention.

本发明的流程如图1所示，具体包括如下的训练步骤①和识别步骤②：The flow process of the present invention is shown in Figure 1, specifically comprises following training step 1. and identification step 2.:

A根据需要按放射源类别——如，标准源或样品的种类及批次等——对已测得的多个能谱进行分类，能谱为在不同测量距离、不同测量时间及其它不同外界条件下所测得；A. Classify the measured multiple energy spectra according to the type of radioactive source—for example, the type and batch of standard source or sample, etc. Measured under the conditions;

B采用小波方法或多项式方法对能谱进行平滑滤波；B adopt wavelet method or polynomial method to smooth and filter the energy spectrum;

C求取能谱的GMM模型(Gaussian mixture model)，即表示为多个高斯分布函数的线性和：C obtains the GMM model (Gaussian mixture model) of the energy spectrum, which is expressed as the linear sum of multiple Gaussian distribution functions:

$P P ((x x,, θ θ)) = = {Σ Σ}_{i i = = 11}^{M m} {a a}_{i i} {p p}_{i i} ((x x;; {θ θ}_{i i})) - - - - - - ((11))$

式(1)中M为高斯分布密度函数的个数，应由能谱的形状、光滑程度、道址数N等而定；a₁，...，a_M是各高斯分布密度函数的权重，即各高斯分布函数在概率密度函数中所占的比重，且

a_i≥0，(i＝1，...，M)，由能谱数据和函数个数M而定；p_i(x)是第i个高斯分布密度函数，其均值μ_i由高斯分布密度函数的个数M和道址数确定；p_i(x)的方差为σ_i ²；θ_i是未知参数μ_i和σ_i ²的向量表示，即

密度函数p_i(x，θ_i)的形式如下：In the formula (1), M is the number of Gaussian distribution density functions, which should be determined by the shape of the energy spectrum, the degree of smoothness, the number of track N, etc.; a ₁ ,..., a _M is the weight of each Gaussian distribution density function , that is, the proportion of each Gaussian distribution function in the probability density function, and

a _i ≥ 0, (i=1,..., M), determined by the energy spectrum data and the number of functions M; p _i (x) is the i-th Gaussian distribution density function, and its mean value μ _i is determined by The number M of density functions and the number of track sites are determined; the variance of p _i (x) is σ _i ² ; θ _i is the vector representation of unknown parameters μ _i and σ _i ² , namely

The form of the density function p _i (x, θ _i ) is as follows:

${p p}_{i i} ((x x,, {θ θ}_{i i})) = = \frac{11}{{((22 π π))}^{11 / / 22} {σ σ}_{i i}} exp exp [[- - \frac{11}{22} {((x x - - {μ μ}_{i i}))}^{22} {(({σ σ}_{i i}^{22}))}^{- - 11}]] - - - - - - ((22))$

整个混合密度的参数为θ＝(a₁，...，a_M；θ₁，...，θ_M)。The parameters of the entire mixture density are θ=(a ₁ , . . . , a _M ; θ ₁ , . . . , θ _M ).

建立GMM模型具体通过以下a、b、c步骤实现：The establishment of the GMM model is achieved through the following steps a, b, and c:

设能谱为F(i)(i＝1...N)，其中N为道址数，M为GMM模型的高斯函数分支数，M通常取N/n(n＝1，2，3...)，总计数为N_total。Let the energy spectrum be F(i) (i=1...N), where N is the number of track sites, M is the number of Gaussian function branches of the GMM model, and M usually takes N/n (n=1, 2, 3. ..), the total count is N _total .

a.将能谱F(i)(i＝1...N)按下式作归一化处理：a. Normalize the energy spectrum F(i) (i=1...N) according to the following formula:

$f f ((i i)) = = F f ((i i)) / / {Σ Σ}_{i i = = 11}^{N N} F f ((i i)),, ((i i = = 11 \cdot &Center Dot; \cdot \cdot \cdot &Center Dot; N N)) - - - - - - ((33))$

b.将归一化后的信号f(i)(i＝1...N)用GMM表示，并得到近似信号f′(i)(i＝1...N)：b. Express the normalized signal f(i) (i=1...N) with GMM, and obtain the approximate signal f'(i)(i=1...N):

${f f}^{' '} ((i i)) = = s the s {Σ Σ}_{j j = = 11}^{M m} f f ((sj sj)) {p p}_{j j} ((i i)) = = {Σ Σ}_{j j = = 11}^{M m} sf sf ((sj sj)) {p p}_{j j} ((i i)),, ((i i = = 11 \cdot \cdot \cdot &Center Dot; \cdot &Center Dot; N N)) - - - - - - ((44))$

其中s＝N/M。where s=N/M.

${p p}_{j j} ((i i)) = = \frac{11}{\sqrt{22 π π} σ σ} exp exp [[- - \frac{11}{22 {σ σ}^{22}} {((i i - - sj sj))}^{22}]],, ((i i = = 11 \cdot \cdot \cdot \cdot \cdot \cdot N N,, j j = = 11 \cdot \cdot \cdot \cdot \cdot \cdot M m)) - - - - - - ((55))$

σ通常取σ＝1～s；σ usually takes σ=1～s;

实际上，可计算得：In fact, it can be calculated that:

${Σ Σ}_{j j = = 11}^{M m} sf sf ((sj sj)) = = 11 - - - - - - ((66))$

(6)式满足(1)式中GMM模型的权值条件：

a_i≥0，(i＝1，...，M)。Formula (6) satisfies the weight condition of the GMM model in formula (1):

a _i ≥ 0, (i=1, . . . , M).

c.按下式计算，并取整后恢复原始能谱：c. Calculate according to the following formula, and restore the original energy spectrum after rounding:

$F f ((i i)) = = {f f}^{' '} ((i i)) \cdot \cdot {N N}_{total total} / / {Σ Σ}_{i i = = 11}^{N N} {f f}^{' '} ((i i)),, ((i i = = 11 \cdot \cdot \cdot \cdot \cdot \cdot N N)) - - - - - - ((77))$

这样，就完成了能谱GMM模型的建立，并将GMM模型中各高斯函数的权值作为降维后的数据；In this way, the establishment of the energy spectrum GMM model is completed, and the weight of each Gaussian function in the GMM model is used as the data after dimension reduction;

D将降维后的能谱数据进行小波包分解，并将各子频带信号的能量进行归一化处理，提取归一化数据作为特征向量，具体通过以下a、b、c步骤实现：D Decompose the energy spectrum data after dimensionality reduction into wavelet packets, normalize the energy of each sub-band signal, and extract the normalized data as feature vectors, specifically through the following steps a, b, and c:

a将降维后的能谱数据进行N层小波包分解，则第N层形成等带宽的2^N个频带，提取从低频到高频各频带的信号分解系数X_j，j＝1，2，...，2^N；a Decompose the energy spectrum data after dimension reduction into N layers of wavelet packets, then the Nth layer forms 2 ^N frequency bands of equal bandwidth, and extracts the signal decomposition coefficient X _j of each frequency band from low frequency to high frequency, j=1, 2, ..., ^2N ;

b由小波包分解系数X_j重构各频带信f_j(t)，则总信号f(t)可表示为b Reconstruct the signal f _j (t) of each frequency band from the wavelet packet decomposition coefficient X _j , then the total signal f (t) can be expressed as

$f f ((t t)) = = {Σ Σ}_{j j = = 11}^{22^{N N}} {f f}_{j j} ((t t)) - - - - - - ((88))$

c求取各频带的能量，并构成特征向量。令第j个频带的能量为E_j，x_jk为重构信号f_j(t)的第k外离散点幅值，则c Calculate the energy of each frequency band and form a feature vector. Let the energy of the j-th frequency band be E _j , and x _jk be the amplitude of the k-th outer discrete point of the reconstructed signal f _j (t), then

${E E.}_{j j} = = {Σ Σ}_{k k = = 11}^{n no} {| | {x x}_{jk jk} | |}^{22},, ((j j = = 1,2 1,2,, . . . . . .,, 22^{N N})) - - - - - - ((99))$

式中n为各频带重构信号f_j(t)的离散点个数；由E_j构成特征向量为X＝[E₁，E₂，…，E_2N]，规一化后变为In the formula, n is the number of discrete points of the reconstructed signal f _j (t) in each frequency band; the eigenvector formed by E _j is X=[E ₁ , E ₂ ,…, E _2N ], which becomes

$X x = = [[\frac{{E E.}_{11}}{E E.},, . . . . . .,, \frac{{E E.}_{22^{N N}}}{E E.}]] - - - - - - ((1010))$

其中 $E = {(Σ_{j = 1}^{2^{N}} {| E_{j} |}^{2})}^{1 / 2};$ in $E. = {(Σ_{j = 1}^{2^{N}} {| {E.}_{j} |}^{2})}^{1 / 2};$

E将特征向量作为样本，采用期望最大化法(Expectation Maximization，简写为EM)对这些样本进行迭代运算，得到各类特征向量的GMM模型，具体通过以下a、b步骤实现：E takes eigenvectors as samples, uses Expectation Maximization (abbreviated as EM) to perform iterative operations on these samples, and obtains the GMM model of various eigenvectors, specifically through the following steps a and b:

a初步建立各类能谱特征向量X的GMM模型：a Preliminary establishment of the GMM model of various energy spectrum feature vectors X:

$P P ((x x,, θ θ)) = = {Σ Σ}_{i i = = 11}^{M m} {a a}_{i i} {p p}_{i i} ((x x;; {θ θ}_{i i})) - - - - - - ((1111))$

其中M是高斯混合密度的混合数，x是维度为d的能谱特征向量，p_i是基本密度，其均值为μ_i，方差矩阵为∑_i，a_i是混合权数。每个基本密度是d维的高斯函数，θ_i是未知参数μ_i和∑_i的向量表示，即θ_i＝(μ_i，∑_i)，密度函数p_i(x，θ_i)的形式如下：Among them, M is the mixture number of Gaussian mixture density, x is the energy spectrum feature vector with dimension d, p _i is the basic density, its mean is μ _i , variance matrix is ∑ _i , and a _i is the mixture weight. Each basic density is a d-dimensional Gaussian function, θ _i is the vector representation of unknown parameters μ _i and ∑ _i , that is, θ _i = (μ _i , ∑ _i ), the form of the density function p _i (x, θ _i ) is as follows :

${p p}_{i i} ((x x,, {θ θ}_{i i})) = = \frac{11}{{((22 π π))}^{d d / / 22} {| | {Σ Σ}_{i i} | |}^{11 / / 22}} exp exp [[- - \frac{11}{22} {((x x - - {μ μ}_{i i}))}^{T T} {(({Σ Σ}_{i i}))}^{- - 11} (({x x - - μ μ}_{i i}))]] - - - - - - ((1212))$

GMM的参数估计采用b步的EM算法；The parameter estimation of GMM adopts the b-step EM algorithm;

b采用期望最大化法，即Expectation Maximization(简写为EM)，将各类特征向量样本代入式(13)～(15)中进行迭代运算直到收敛，实现高斯混合模型参数的更新并最终得到各类样本的GMM模型；bUsing the expectation maximization method, that is, Expectation Maximization (abbreviated as EM), various types of eigenvector samples are substituted into formulas (13)~(15) for iterative operations until convergence, so as to update the parameters of the Gaussian mixture model and finally obtain various The GMM model of the sample;

${a a}_{l l}^{new new} = = \frac{11}{N N} {Σ Σ}_{i i = = 11}^{N N} p p ((l l | | {x x}_{i i},, {θ θ}^{g g})) - - - - - - ((1313))$

${μ μ}_{l l}^{new new} = = \frac{{Σ Σ}_{i i = = 11}^{N N} {x x}_{i i} p p ((l l | | {x x}_{i i},, {θ θ}^{g g}))}{{Σ Σ}_{i i = = 11}^{N N} p p ((l l | | {x x}_{i i},, {θ θ}^{g g}))} - - - - - - ((1414))$

${σ σ}_{l l}^{22 ((new new))} = = \frac{{Σ Σ}_{i i = = 11}^{N N} p p ((l l | | {x x}_{i i},, {θ θ}^{g g})) {(({x x}_{i i} - - {μ μ}_{l l}^{new new}))}^{22}}{{Σ Σ}_{i i = = 11}^{N N} p p ((l l | | {x x}_{i i},, {θ θ}^{g g}))} - - - - - - ((1515))$

式(13)～(15)中N为各类特征向量样本的数目，x_i为特征向量样本，a_l、μ_l及σ_l ²分别是各高斯分布函数的权重、均值及方差；In formulas (13)-(15), N is the number of various feature vector samples, x _i is the feature vector sample, a _l , μ _l and σ _l ² are the weight, mean and variance of each Gaussian distribution function respectively;

通过以上A～E步骤步即完成能谱的训练过程；Through the above steps A to E, the training process of the energy spectrum is completed step by step;

B求取待识别能谱的GMM模型(Gaussian mixture model)，并将GMM模型中各高斯函数的权值作为降维后的数据，方法同①步骤中的C步骤；B Find the GMM model (Gaussian mixture model) of the energy spectrum to be identified, and use the weights of each Gaussian function in the GMM model as the data after dimension reduction, the method is the same as step C in step ①;

C将降维后的待识别能谱数据进行小波包分解，并将各子频带信号的能量进行归一化处理，提取归一化数据作为特征向量，方法同①步骤中的D步骤；C decomposes the energy spectrum data to be identified after dimension reduction by wavelet packet, and normalizes the energy of each sub-band signal, and extracts the normalized data as a feature vector, the method is the same as step D in step ①;

D将待识别能谱的特征向量作为多维随机数，并计算其分属于各类GMM模型的类条件概率密度，方法如下：D uses the eigenvector of the energy spectrum to be identified as a multidimensional random number, and calculates the class conditional probability density belonging to each type of GMM model, the method is as follows:

选取待识别能谱的特征向量x＝{x₁，x₂，...x_d}，x₁，x₂，...，x_d为d种特征量，其类条件概率密度为Select the feature vector x={x ₁ , x ₂ ,...x _d } of the energy spectrum to be identified, x ₁ , x ₂ ,..., x _d are d kinds of feature quantities, and the class conditional probability density is

$p p ((x x | | {C C}_{i i})) = = {Π Π}_{n no = = 11}^{d d} p p (({x x}_{n no} | | {C C}_{i i})),, i i = = 11,, . . . . . .,, m m - - - - - - ((1616))$

其中p(x_n|C_i)为C_i类的第n个特征的类条件概率密度函数，m为类别数；Where p(x _n |C _i ) is the class conditional probability density function of the nth feature of class C _i , and m is the number of categories;

E最后按贝叶斯决策分类，判决过程如下a、b步骤：E is finally classified according to Bayesian decision-making, and the judgment process is as follows in steps a and b:

a首先，计算 $Σ_{n = 1}^{d} \log (p (x_{n} | C_{i})), i = 1, . . ., m .$ aFirst, calculate $Σ_{no = 1}^{d} \log (p (x_{no} | C_{i})), i = 1, . . ., m .$

b如果

则判断x{x₁，x₂，...，x_n}∈C_j；bif

Then judge x{x ₁ , x ₂ ,..., x _n }∈C _j ;

通过以上A～E步骤步即完成待识别能谱的识别过程。Through the above steps A to E, the identification process of the energy spectrum to be identified is completed step by step.

从上述的放射性能谱识别方法可以看出，本发明结合了放射性测量中能谱的统计特性及非平稳特性，在训练阶段先采用小波方法或多项式方法对能谱进行平滑滤波，以消除干扰信号；接着用GMM模型对能谱作降维处理，以保证能谱的统计特性在低维空间仍具有原始能谱信号的绝大部分信息；然后，在多频带信号空间提取归一化特征向量并作为样本，保证了采用低维样本代表能谱特征的唯一性，同时也保证了样本与测量时间的无关性；最后，采用期望最大化法对这些样本进行迭代运算，得到各类特征向量的GMM模型，这样既保证了准确收敛又可通过合理调整GMM模型的高斯函数个数以满足各种不同场合下识别的需要。另外，本发明在识别时，采用贝叶斯决策分类，保证了识别准确率在统计意义下的最优。As can be seen from the above radioactivity spectrum identification method, the present invention combines the statistical properties and non-stationary properties of the energy spectrum in radioactivity measurement, and first adopts wavelet method or polynomial method to smooth and filter the energy spectrum in the training stage to eliminate interference signals ; Then use the GMM model to reduce the dimensionality of the energy spectrum to ensure that the statistical properties of the energy spectrum still have most of the information of the original energy spectrum signal in the low-dimensional space; then, extract the normalized feature vector in the multi-band signal space and As a sample, the uniqueness of using low-dimensional samples to represent the energy spectrum features is guaranteed, and the independence of samples and measurement time is also guaranteed; finally, the expectation maximization method is used to iteratively calculate these samples to obtain the GMM of various feature vectors In this way, accurate convergence can be ensured and the number of Gaussian functions of the GMM model can be adjusted reasonably to meet the needs of recognition in various situations. In addition, the present invention adopts Bayesian decision-making classification during recognition, which ensures the optimal recognition accuracy rate in a statistical sense.

总之，本发明是一种既能快速训练又能避免不能收敛或局部收敛的方法，使能谱的识别能顺利进行且能保证更高的识别准确率。In a word, the present invention is a method capable of fast training and avoiding non-convergence or partial convergence, so that the identification of energy spectrum can be carried out smoothly and can ensure higher identification accuracy.

在上述本发明的实施例中，对放射性能谱识别方法进行了详细说明，但需说明的是，以上所述仅为本发明的一个实施例而已，本发明同样可对局部谱段的能谱进行识别，可用于各种射线能谱的识别，凡在本发明的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。In the above-mentioned embodiments of the present invention, the radioactive energy spectrum identification method has been described in detail, but it should be noted that the above is only an embodiment of the present invention, and the present invention can also analyze the energy spectrum of the local spectrum The identification can be used for identification of various ray energy spectra. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A radioactivity spectrum identification method, characterized in that, the specific steps are as follows:

① Training on the energy spectrum measured in radioactivity measurement;

② Decision-making classification of the energy spectrum to be identified.

2. A kind of radioactive energy spectrum identification method according to claim 1, is characterized in that, in described 1., the energy spectrum measured in the radioactive measurement is trained, comprising the following steps:

A Filter, dimensionally reduce and classify the measured energy spectrum in radioactivity measurement,

B extracts the eigenvectors of the processed energy spectrum,

C takes the feature vector as a sample, and obtains the GMM model of various feature vectors through training.

3. A kind of radioactive energy spectrum identification method according to claim 2, it is characterized in that, in said A: filtering refers to adopting wavelet method or polynomial method to carry out smooth filtering to energy spectrum; Dimensionality reduction refers to seeking energy spectrum The GMM model (Gaussian mixture model), and the weight of each Gaussian function in the GMM model is used as the data after dimension reduction; classification refers to classification by radioactive source.

4. a kind of radioactivity spectrum identification method according to claim 2, it is characterized in that, extracting feature vector in described B refers to carrying out wavelet packet decomposition with the data after claim 3 dimensionality reduction, and each sub-frequency band The energy of the signal is normalized, and the normalized data is extracted as a feature vector.

5. a kind of radioactivity spectrum identification method according to claim 2, is characterized in that, obtains the GMM model of various feature vectors by training in the described C, refers to the feature vector that claim 4 extracts as multidimensional random number , using the Expectation Maximization method (EM for short) to perform iterative operations on these random numbers to obtain the GMM model of various feature vectors.

6. A kind of radioactive energy spectrum identification method according to claim 1, is characterized in that, in described ②, energy spectrum to be identified is carried out decision-making classification, comprises the following steps:

aUse the wavelet method or polynomial method to smooth and filter the energy spectrum to be identified, and obtain the eigenvector of the energy spectrum to be identified,

b Obtain the class conditional probability density of the feature vector according to the GMM model,

c Then classify by Bayesian decision.

7. A kind of radioactive energy spectrum identification method according to claim 6, it is characterized in that, in the described a, seek the eigenvector of energy spectrum to be identified, refer to energy spectrum to be identified and carry out after dimensionality reduction according to claim 3 The wavelet packet is decomposed, and the energy of each sub-band signal is normalized, and the normalized data is extracted as a feature vector.

8. a kind of radioactivity spectrum identification method according to claim 6, is characterized in that, obtains the class conditional probability density of feature vector by GMM model among the described b, refers to the energy spectrum to be identified that claim 7 extracts The eigenvectors are used as multidimensional random numbers to be classified, and the class conditional probability densities belonging to various GMM models are calculated.